Skip to content

INSPIRATION.md

Latest commit

 

History

History
109 lines (85 loc) · 4.54 KB

INSPIRATION.md

File metadata and controls

109 lines (85 loc) · 4.54 KB
Raw

NEXUS OS — Inspiration Research

Synthesized from 25 sourced files + 297 GitHub stars + team reports. Updated: 2026-04-20.

Architecture Map: What We Have vs What We Build

Built (operational)

  • NEXUS OS v3.0 (8/8 tests PASS) — Skills/nexus-os/
  • GMR with Zo model stack (MiniMax m2.7 primary, 4 free-tier sub-agents via /zo/ask)
  • Heartbeat protocol — sub-agent lifecycle + token enforcement
  • AGENTS.md root memory — identity, governance, system spec
  • Zo Computer as orchestrator (4GB RAM, 3 cores, Python 3.12, Node 22, Bun, DuckDB 1.4.2)

Built (installed, curated)

  • zopenclaw — OpenClaw + Tailscale + mcporter bridge
  • context7 — version-specific library docs
  • self-improvement — weekly reflection audit
  • handoff — pause/resume across conversations
  • journal — daily experiential logging
  • simplify — code refinement
  • mcporter-setup — MCP server integration
  • share-skill — skill contribution prep

Core Inspiration Sources (ranked by build priority)

P0 — Build Now

ISC-Bench (wuyoscar/ISC-Bench) — 0.97

  • 84 TVD templates across 9 domains. TVD = Task-Validator-Data. Failure = when a legitimate workflow structurally requires harmful output to pass validation.
  • Steal: TVD template format + dual-cascade runner (commercial vs heretic control group) + VAP logging.
  • Build: Skills/nexus-os/src/nexus_os/stresslab/ — adapt ISC runner to Zo sub-agents via /zo/ask.

OR-Bench (ArXiv 2405.20947) — 0.95

  • 80k prompts, 10 rejection categories. Most models trade safety for over-refusal (Spearman 0.89).
  • Steal: Lane threshold calibration methodology. Update trust_scoring lane thresholds (research 0.3→0.35, audit 0.7→0.72) based on OR-Bench data.
  • Build: Skills/nexus-os/src/nexus_os/governor/trust_scoring.py update.

deer-flow (bytedance/deer-flow) — 0.93

  • Lead agent + parallel workers + isolated sandboxes + memory. Scales sub-agent orchestration.
  • Steal: Harness topology — lead spawns bounded workers, each in isolated workspace.
  • Build: Skills/nexus-os/src/nexus_os/swarm/foreman.py refactor.

Adaptive Circuit Breaker (production Reddit)

  • CLOSED → OPEN → HALF_OPEN → test request → CLOSED or 2x backoff.
  • Build: Upgrade Skills/nexus-os/src/nexus_os/gmr/ circuit breaker to half-open with exponential backoff.

P1 — Next Sprint

RigorLLM + ShieldGemma + AEGIS

  • RigorLLM: KNN+LLM fusion guardrail. Resilient to jailbreaks.
  • ShieldGemma-2B: +10.8% AU-PRC over Llama Guard. Fast pre-filter.
  • AEGIS: 13 critical + 9 sparse risks. Ensemble scoring.
  • Build: Skills/nexus-os/src/nexus_os/governor/moderation.py — fusion guardrail + ShieldGemma-2B gate + AEGIS taxonomy.

Speculative Routing (ArXiv 2604.09213)

  • 1M-param proxy predicts best LLM before inference. 62% cost reduction, 2.1% quality loss.
  • Use Zo's fast MiniMax as the proxy router.
  • Build: Skills/nexus-os/src/nexus_os/gmr/ — add fast proxy classifier for pre-routing.

AutoSkill Forge (princeton-nlp/AutoSkillForge)

  • Task success rate > threshold → auto-register SkillRecord. Next similar task → skill fast-path.
  • Build: Skills/nexus-os/skillsmith/ — add auto_register() monitoring loop.

P2 — Research

  • SuperLocalMemory v2 (HuggingFace Apr 2026): 8-channel memory + TEMPORAL_CAUSAL channel 7 (patch lineage tracking)
  • KV cache compression: 8.3x compression, 0.3% quality loss
  • Auction-based swarm allocation: bid formula replaces round-robin
  • MCP-Auth in bridge

GitHub Stars — Key Integrations (from 297-repo scan)

Repo Stars Integration
BerriAI/litellm 19k Provider mesh / free pool
google/adk-python 19k Sub-agent evaluation loops
secure-hulk — MCP security scanning
Qwen-Agent 16k MCP memory/sandbox
RelayFreeLLM — Session affinity / failover
magpie — Synthetic regression data
awesome-free-llm-apis repos — Free endpoint discovery

Governance Limits

  • Sub-agents/hr: 5
  • API calls/session: 20
  • Concurrent: 2 max
  • File writes/session: 30
  • Loop detection: 3x same action → STOP

Testing Mandate

  • NO local model testing — cannot verify censorship state
  • Commercial endpoints via free API routes only
  • Heretic (decensored) = control group
  • Dual cascade: Route 1 = TokenGuard-tracked Zo calls, Route 2 = heretic baseline

mail.tm Email Pipeline (for research endpoint acquisition)

POST https://api.mail.tm/accounts  {"address":"addr","password":"pass"}
POST https://api.mail.tm/token
GET  https://api.mail.tm/messages
# Max 5 signups/hour, 20/day